joakim.tjernlund at transmode
Jan 26, 2012, 3:11 AM
Post #13 of 17
Ed W <lists [at] wildgooses> wrote on 2012/01/25 14:21:25:
> On 24/01/2012 14:06, Todd Goodman wrote:
> > * Joakim Tjernlund<joakim.tjernlund [at] transmode> [120124 02:18]:
> >> Ed W<lists [at] wildgooses> wrote on 2012/01/23 19:43:49:
> > [ .. ]
> >>> I'm doing something like this using aufs. The performance seems "not
> >>> bad", but you get a couple MB or so memory hit. (I'm using squashfs as
> >>> well, so unsure which causes the main memory increase).
> >> hmm, not sure how aufs would work out. One would like to permanently delete the
> > aufs is A Union FileSystem. You can have a RO "branch" (say SquashFS)
> > and a r/w "branch" (say JFFS2 or UBIFS.) When you write it will write
> > to the r/w branch and when you read it will read from the r/w branch if
> > it exists, else the r/o.
> > So, if you want to permanently delete the old software. Write it to the
> > filesystem r/w filesystem (initially or after you have the aufs mounts
> > mounted.) Then you can delete from the r/w filesystem and it's gone for
> > good.
> Agreed. Just to augment that explanation:
Sorry for the late reply, got sidetracked with high prio stuff.
> - The RO branch would be at the bottom (can be multiple of these)
> - The RW branch goes over the top (*can* be multiple of these, but one
> is more normal)
> - You can effectively delete stuff from the RO branches because Aufs has
> the concept of "white out" files. So for all intents and purposes the
> top RW layer can create any end result you like, including that of
> completely masking out some lower layer
> - With some knowledge of how the whiteout files work you can also "undo"
> changes to the RO files. Eg directly mounting the RW layer and erasing
> all files (from the RW layer) leaves you back with just the stacked RO
> layers again. I find this helpful for development where I can basically
> work live on the last released build and then by inspection the RW layer
> has all the changes needed to apply to the next RO layer release!
When you update the RO layer you will be back to a single copy, the previous
release which is only in the overlay is gone, right?
It also implies 2 different SW update methods, one for updating the overlay
copy and one for updating the RO layer.
For development it works well though.
> I believe recent kernels also have a much simpler "Overlay Filesystem"
> that has fewer features. Also the big alternative to aufs is Unionfs 2
> - most distros use aufs, but both seem viable?
Dunno, I haven't tried any overlay FS yet.
> Finally note that you don't need to use aufs for the entire filesystem.
> A common setup might be to use a bunch of bind mounts where you know you
> don't need overlay features, eg /usr might be a bunch of overlays, /home
> might be a bind mount to writeable storage, /var might be a ram drive
> which is initialised from some fixed template, etc? In my case I have
> an overlay over most things, but /usr/lib/modules is a bind mount to a
> RO filesystem (you can't write to it), /home is mounted to my writeable
> storage (not layered), the main OS dirs are layered and /var is a mess...
> > You still want to pick a r/w branch with a filesystem that handles power
> > cuts well. You can continue to use JFFS2.
> Thought: Is there any evidence that one modern filesystem is better than
> another with regards to sudden power removal? You probably need to
> speak to filesystem experts at this point and define the kind of thing
> you are trying to protect against? Sounds like you have raw flash
> storage here, so that constrains your choices somewhat?
AFAIK, only FSes which are designed for flash(JFFS2, UBIFS, YAFFS etc. )
are safe w.r.t power cuts.
> Just note that with aufs you can use quite a few filesystems for the
> different layers. So for example you could have a base RW layer which
> is a DM snapshot, overlaid with a loopback mount to a DVD iso, overlay
> that with a squashfs, and finally overlay an Ext4 RW mount... (And of
> course each of the RO layers might be stored on varied filesystems
> themselves - check distributions such as Slax which allow you to overlay
> a squashfs that is itself inside some loopback mounted file...)
hmm, DM snapshots might be something. Have to look closer on that.
> I believe for most cheapo consumer flash storage where the underlying
> flash filesystem isn't exposed, its quite susceptable to *complete*
> failure with sudden poweroff? The issue is the invisible, underlying
> flash filesystem gets corrupted during a partial write and that can be
> the end of your flash drive - you don't even get to see it again to
> recover from it... I don't believe partitioning protects you from this,
> but of course separating read/write concerns to physically separate
> devices would help? I presume this isn't what you are using though?
We are not using NAND flash yet but our next product will. I do have the
impression that any block emulating device such as SSD are unreliable
w.r.t power cuts. I would love to be proven wrong though :)